Achieving LanguageInteroperability with

Gary Kumfert with

Bill Bosl, Tammy Dahlgren, Tom Epperly, Scott Kohn, &

Steve Smith

GKK 2CASC

Babel’s Scope & CCA� Babel provides language

interoperability, not components.

� We collaborate with CCA to add parallel distributed support

� We also provide tools (Quorum & Alexandria) to facilitate component development and

deployment

Compilers &Linkers

Operating System

CCA Compliant FrameworksComponentSemantics

GKK 3CASC

Release Announcement

� C, C++, F77, Python(client)

GKK 4CASC

babel-0.5.0.tar.gz

� Babel code generator� written in Java

� Babel runtime library� written in ANSI C

� Docs (minimal)� papers, talks, javadoc html� babel101 tutorial

GKK 5CASC

Hand Coded Language Interoperability

JNINativeSWIG

Platform Dependent

C

C++

f77

f90

Python

Java

GKK 6CASC

Babel Enabled Language Interoperability

� Truly Object Oriented� Reference Counting

� Exception Handling� RMI (future)

C

C++

f77

f90

Python

Java

GKK 7CASC

Babel Enabled Language Interoperability

What’s In This Release:What’s In This Release:

C

C++

f77

f90

Python

Java

GKK 8CASC

Babel Has Two Types of Customers

� Have a code� Want to increase

their user base� Will learn SIDL� Want Babel

general and powerful

� Have a problem� Want to solve their

problem� May never see SIDL � Want software that’s

easy & trustworthy

Developers Users

GKK 9CASC

Babel’s Design Priorities

� Performance� Developer/User dichotomy� What’s natural for each language?

� Could expose C array structs in C++� C++ style would be SIDL::array<T>template <>array<item_t> : public array_mixin

< array_t, item_t, item_cxx_wrapper_t> { }

template <>array<item_t> : public array_mixin

< array_t, item_t, item_cxx_wrapper_t> { }

GKK 10CASC

SIDL (Scientific Interface Definition Language)

� Builds on Industry IDL technology

� CORBA� COM

� Designed for Scientific Apps

� complex types� dynamic

multidimensional arrays

version Hello 1.0;package Hello { class World {string getMsg();

}}

GKK 11CASC

version MySolverLib 0.1.0;import ESI; package MySolverLib { interface MatrixGenerator { ... } class OptionDatabase {void getOption( in string name,

out string val);}class Vector implements-all ESI.Vector {void setOptions( in OptionDatabase db );

}class Bizarre implements MatrixGenerator {

... void setData( in array<dcomplex,2> a );

}}

GKK 12CASC

Many forms of language interoperable interfaces

SIDLScientific Interface

Definition Language

Human Comp

IORInternal ObjectRepresentation

mpiCo

XMLeXtensible Markup

Languageb Compatible

atible

We

ler Compatible

GKK 13CASC

XML enables......automated creationvia higher-level tools...Type Descriptions

on Shared Repositories

...Browsingfor Types

...automated search &

discovery by advanced

builders

GKK 14CASC

Language Interoperability: How Babel Makes it Work

Impls

� Application: user’s LOP(Language Of Preference)

� Implementation: developers LOP (Can be wrappers to legacy code)

SIDLSIDL

If developer used Babel,they also have a SIDL file.

Stubs

Skels

IORs

Application

GKK 15CASC

Language Interoperability: How Babel Makes it Work

Application

Stubs

Skels

IORs

Impls

� Application: user’s LOP(Language Of Preference)

� Implementation: developers LOP (Can be wrappers to legacy code)

� Client Side Stubs: user’s LOP to C

� Internal Object Representation (IOR): Always in C

� Server Side Skeletons: translates IOR (in C) to developer’s LOP

GKK 16CASC

How Much Language Interoperability Have We Achieved?

� Basic Types� bool� char� int� long� float � double� fcomplex� dcomplex� string� opaque

� Basic Types� bool� char� int� long� float � double� fcomplex� dcomplex� string� opaque

� Extended Types� Objects� enumerations� arrays of basic types� arrays of objects� (arrays are

multidimensional, no arrays of arrays)

� Extended Types� Objects� enumerations� arrays of basic types� arrays of objects� (arrays are

multidimensional, no arrays of arrays)

� Modes� in� out� inout� return

value

� Modes� in� out� inout� return

value

� OO Method Dispatch� regular� final� static� interfaces� classes

� OO Method Dispatch� regular� final� static� interfaces� classes

� Exception Handling� Exception Handling

� For All Combinations of Languages� C C� C++ C++� F77 F77� Python

� For All Combinations of Languages� C C� C++ C++� F77 F77� Python

� 1431 test cases (and counting)

� 1431 test cases (and counting)

GKK 17CASC

0

200

400

600

800

1000

1200

1400

1600

Jan-00

Feb-00

Mar-00

Apr-00

May-00Jun-00Jul-0

0Aug-00Sep

-00Oct-

00Nov-00Dec-00Jan

-01Feb

-01Mar-

01Apr-0

1May-01Jun-01Jul-0

1

passxfailfail

Test Historysparc-sun-solaris2.7-gcc#

Test

Cas

es

Date

babel 0.5.x

babel 0.4.x

C++ start

Parser start

ANSI C start

Python start

F77 start

IOR

Rew

rite

GKK 18CASC

What we foresee, based on experience with our tests...

Language Interoperability

Developer Concerns:Configuration, Packaging,& Deployment

User Concerns:InstallationTrust

GKK 19CASC

GKK 4CASC

Hand CodedLanguage Interoperability

JNINativeSWIG

Platform Dependent

C

C++

f77

f90

Python

JavaGKK 4CASC

Hand CodedLanguage Interoperability

JNINativeSWIG

Platform Dependent

C

C++

f77

f90

Python

Java

These aren’t new problems...But they are on a larger scale

Developer Concerns:Configuration, Packaging,& Deployment

User Concerns:InstallationTrust

Babel enabledLanguage Interoperability

� Object Oriented� Reference Counting

� Exception Handling� RMI (future)

C

C++

f77

f90

Python

JavaGKK 5CASC

Babel enabledLanguage Interoperability

� Object Oriented� Reference Counting

� Exception Handling� RMI (future)

C

C++

f77

f90

Python

Java

GKK 5CASC

GKK 20CASC

For Example...� Consider the following statements...

� Java is more portable than C� C is more portable than C++ � C++ is more portable than F77

� “portable” in a different sense� Java doesn’t need to be recompiled (like C)� C is mature, isn’t as hard to parse (as C++)� C++ can bind to C or Java easier than F77

� BUT F77 is also more portable than C because of header/library issues

Babel’s

regression tests

have the worst of

all worlds!!!

GKK 21CASC

For Example...� To support Python and Java

� All libraries must be shared (*.so) not statically linked (*.a)

� C++ shared libraries are problematic � Exception support is platform/compiler

dependent� Linking issues when interoperating with

other languages� Can create valid shared library with

uncatchable exceptions

GKK 22CASC

Babel’s Configuration/Build� GNU Make � Autoconf

configuration� Automake

build Makefiles� Libtool

shared libraries� CUTE

custom testing

� python’s own build system

� java’s built-in (broken) make

� helper scripts� fixes to autoconf,

automake &libtool

� lots of hacks

GKK 23CASC

0

200

400

600

800

1000

1200

1400

1600

Jan-00

Feb-00Mar-

00Apr-0

0May-00Jun-00Jul-0

0Aug-00Sep-00Oct-0

0Nov-00Dec-00Jan

-01Feb-01Mar-

01Apr-0

1May-01Jun-01Jul-0

1

passxfailfail

Test Historysparc-sun-solaris2.7-gcc#

Test

Cas

es

Date

babel 0.5.x

babel 0.4.x

IOR

Rew

rite

Configure/Build/Repository faults,not software failures!!!

GKK 24CASC

Problems affect the User too...� How does a user get and install

“language interoperable” software?� Binary: if supplied by developer� Source:

Assume “configure; make install”?� How to link into application?

� If any C++ code, must use C++ linker� Which C++ to use?

C++ has no std binary interface

GKK 25CASC

Crux of the problem

� We’re building 21st century technology...

� ... using 30 year old tools.

MakeBourne shell

Autoconf (M4)Automake (perl)libtool (perl/sh)

JavaPythonF77C++C

Autoheaderaclocal

GKK 26CASC

Solution

� Integrated config, build, package, test and management system.� no make inside!

can have action create many filesunderstands directories

� uses real database� program all aspects in one language� MUST BE OPEN SOURCE

GKK 27CASC

In the mean time...

� Babel works on other platforms, just not automated config/testing� Java code generator (precompiled)� ANSI C runtime library (no problem)

� Babel’s tests are (necessarily) pathological worst-case examples

� We didn’t create these problems, we just exercise them aggressively

GKK 28CASC

Future Babel: Will Provide More Build Help

� “babel.make” � currently lists code generated� may add additional flags, macros, etc.

� configure� currently used for regression tests� may generate artifacts useful for

developershelper scriptswarnings

GKK 29CASC

Farther Future Babel:Will Do Distributed Computing

Internet

Marshaler

Line Protocol

Line Protocol

UnmarshalerStubs

IORs

Skels

IORs

Application

Impls

GKK 30CASC

Closing Remarks

� Babel Beta 0.5 is released� Babel enables language

interoperability � connect C, C++, F77, and Python � provide a uniform object-model,

even in non-OO languages. � Deploying & Installing Language

Interoperable Code in General� is still very hard� has broken every tool we use

GKK 31CASC

The Endbabel-announce@llnl.gov

babel-users@llnl.gov

http://www.llnl.gov/CASC/componentscomponents@llnl.gov

Bill Bosl, Tammy Dahlgren, Tom Epperly, Scott Kohn,

Gary Kumfert, & Steve Smith

GKK 32CASC

A.3. Can HPC and Component Technology REALISTICALLY be integrated?

� Yes.� But HPC Components have huge

(and unique) hurdles:� Diverse Architectures� Diverse OS’s� Integration of SPMD and Dist. Comp.� Archaic Pkg/Devel/Config/Build tools� Non-CS trained (or interested) users

GKK 33CASC

B.3. Can the HPC community really afford yet another compiler such asBabel?

� Is language interoperability important?

� How important?

GKK 34CASC

B.3. How is the Java subset of C++ inadequate as an HPC IDL?

� What is a “Java subset of C++” ?� How does one

� use it to bind to other languages? � get a common inheritance model?� get a common exception model?

GKK 35CASC

B.5. What is the role of traditional (parallel)tools in component technology?

� Hopefully, they’re replaced by modern parallel tools.

GKK 36CASC

B.9. What will be the configuration issuesfor components...to be portable andhigh-performance?

� Lots.� Lack of Configuration, Packaging,

& Deployment tools � is the #1 Achilles heel for

components� is the #1 day-to-day pain in Babel

development� #1 cause for failure in regression

tests

GKK 37CASC

C.1. Will anyone actually make the effort tocomponentize their applicationssoftware?

� Yes.� But it will be messy.

� Efforts to Babelize at LLNL:� hypre - want OOP in ANSI C &

automatic F77 bindings� ALPS - want scripting interface for

laser plasma physics � SAMRAI - framework used in ALPS

GKK 38CASC

D.4. Should components be viewed as mostly a library/runtime developertechnology?

� No. � I used components in this

PowerPoint Presentation� Users can use components without

knowing they’re using them.� This is harder to achieve in UNIX than

other platforms

GKK 39CASC

23 Jul 2001UCRL-PRES-144649

Work performed under the auspices of the U. S. Department of Energy by the University of California, Lawrence Livermore National Laboratory under Contract W-7405-Eng-48